Although instruction tuning is widely used to adjust behavior in Large Language Models (LLMs), extensive empirical evidence and research indicates that it is primarily a process where the model fits to specific task formats, rather than acquiring new knowledge or capabilities. We propose that this limitation stems from biased features learned during instruction tuning, which differ from ideal task-specfic features, leading to learn less underlying semantics in downstream tasks. However, ideal features are unknown and incalculable, constraining past work to rely on prior knowledge to assist reasoning or training, which limits LLMs' capabilities to the developers' abilities, rather than data-driven scalable learning. In our paper, through our novel data synthesis method, DELIA (Diversity-Enhanced Learning for Instruction Adaptation), we leverage the buffering effect of extensive diverse data in LLMs training to transform biased features in instruction tuning into approximations of ideal features, without explicit prior ideal features. Experiments show DELIA's better performance compared to common instruction tuning and other baselines. It outperforms common instruction tuning by 17.07%-33.41% on Icelandic-English translation bleurt score (WMT-21 dataset, gemma-7b-it) and improves accuracy by 36.1% on formatted text generation (Llama2-7b-chat). Notably, among knowledge injection methods we've known, DELIA uniquely align the internal representations of new special tokens with their prior semantics.

Feature selection measures are often explained by the analogy to a rule to measure the "distance" of sets of features to the "closest" ideal sets of features. An ideal feature set is such that it can determine classes uniquely and correctly. This way of explanation was just an analogy before this paper. In this paper, we show a way to map arbitrary feature sets of datasets into a common metric space, which is indexed by a real number p with 1 p . Since this determines the distance between an arbitrary pair of feature sets, even if they belong to different datasets, the distance of a feature set to the closest ideal feature set can be used as a feature selection measure. Surprisingly, when p 1, the measure is identical to the Bayesian risk, which is probably the feature selection measure that is used the most widely in the literature. For 1 p, the measure is novel and has significantly different properties from the Bayesian risk. We also investigate the correlation between measurements by these measures and classification accuracy through experiments. As a result, we show that our novel measures with p 1 exhibit stronger correlation than the Bayesian risk.